Reproduce “DiffuMural: Restoring Dunhuang Murals with Multi-scale Diffusion” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Puyu Han, Jiaju Kang, Yuhang Pan, Erting Pan, Zeyu Zhang, Qunchao Jin, Juntao Jiang, Zhichen Liu and 1 more

DiffuMural, developed by researchers from Southern University of Science and Technology, Beijing Normal University, and others, introduces a multi-scale diffusion framework for high-resolution restoration of large-area missing Dunhuang murals by leveraging contour-guided conditioning and frequency-domain optimization. The framework demonstrated superior structural similarity (up to 45.45% improvement) and edge consistency (up to 28.24% improvement) over other methods, alongside high expert evaluation scores for cultural and artistic authenticity.

Key insights

  • Integrating multi-scale diffusion models with adaptively weighted influence functions effectively fuses global structure and local details for large-area restoration, even with limited, stylistically consistent training data.
  • Leveraging subtle 'damage contour' information as explicit conditional guidance significantly improves the accuracy and stylistic consistency of generated content compared to methods lacking such specific cues.
  • A comprehensive evaluation framework combining objective quantitative metrics (SSIM, CCON, TCON, ECON) with subjective expert assessment on humanistic values is crucial for validating the artistic authenticity and cultural consistency in digital heritage restoration.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.